How (not) to get hit by a self-driving car

A street-based game that challenges visitors to avoid being detected in the eye of an AI.

2023 — Co-created with Daniel Coppen (Playfool)

What if you could deliberately go and collide with self-driving cars to reveal their blind spots?

With advanced image recognition systems becoming increasingly prevalent through surveillance cameras and self-driving cars, the cities we live in are starting to see us back. But how do these systems observe us, and where do their blind spots lie? This game welcomes anyone to challenge the AI and try to reach the goal without being detected. AI in self-driving car has a challenge known as "edge cases," where unexpected events can lead to errors.

For instance, if a child completely covered in cardboard suddenly appears in front of it, the AI cannot decide whether to proceed because it's just cardboard, or whether to apply emergency brakes considering the potential risk of a child inside, even if it puts the driver at risk. To win the game, players need to do unconventional things, such as tumbling or wearing cones. In other words, winning the game results in generating edge cases where the AI couldn't recognize a pedestrian.  Data from these edge cases can potentially complement the flaws in the algorithms or datasets that the AI was based on, possibly contributing to the future performance improvement of self-driving cars.

「自動運転車とぶつかる/ぶつからない方法」は、AIに「歩行者」と検知されないように横断歩道を渡り切る、路上を舞台にしたゲームです。このゲームに勝つためには、AIの目を巧みに騙し、歩行者と検知されることなくゴールに辿り着くユニークな方法を見つけ出す必要があります。自動運転のAIには、「エッジケース」と呼ばれる想定外の出来事に遭遇した際にエラーを起こしてしまうという課題があります。例えば、ダンボールをすっぽり被った子供が目の前に現れた時に、ダンボールだから直進して轢いてもいいのか、それとも子供が中にいる可能性があるから運転手を危険に晒してでも急ブレーキを踏むべきなのかといった判断ができません。このゲームでプレイヤーは勝つためには、側転をしたりコーンを被ったりと奇抜なことをする必要があります。つまり、ゲームに勝つことが、AIが歩行者を認識できなかった場合のエッジケースを生成することになります。これらのエッジケースのデータはAIの元となったアルゴリズムやデータセットの欠陥を補完しうるデータとなりうるため、将来的に自動運転車の性能向上に寄与する可能性があります。

Every player’s win generates data that exposes the inability of the system to detect pedestrians and highlights the flaws of these algorithms

When they reach the goal, players are presented with a final dilemma upon victory: either train the AI or not. They can opt-in their anonymised gameplay footage to improve the AI models, or immediately delete the image. This poses the question of whether people are willing to trade their data for a potentially safer system, or if they would rather remain invisible, despite the implied risks of inaccurate systems in the future.

Project Details

Commissioned project Playable City 2023

Credit

Collaborator » Dan Coppen (Playfool)
Sound Design » Plot Generica
Photography » Luke O'Donovan